An Offline Mobile Conversational Agent for Mental Health Support: Learning from Emotional Dialogues and Psychological Texts with Student-Centered Evaluation

TL;DR

The paper presents EmoSApp, a fully offline smartphone-based mental health assistant built on a fine-tuned and quantized LLaMA-3.2-1B-Instruct, enabling private, on-device inference with modest RAM. It fuses a large Knowledge Dataset (14,582 QA pairs) with ESConv and ServeForEmo to achieve domain-specific empathy and guidance, exploring full fine-tuning, LoRA+PTQ, and QAT-LoRA approaches. Quantitative benchmarks show a trade-off between accuracy and efficiency, with QAT-LoRA offering a strong on-device balance, while qualitative studies with students and professionals demonstrate improved empathy and contextual relevance. The work demonstrates the feasibility and practical value of private, portable AI-driven mental health support for students, while acknowledging limitations and outlining paths for localization and safety enhancements.

Abstract

Mental health plays a crucial role in the overall well-being of an individual. In recent years, digital platforms have increasingly been used to expand mental health and emotional support. However, there are persistent challenges related to limited user accessibility, internet connectivity, and data privacy, which highlight the need for an offline, smartphone-based solutions. To address these challenges, we propose EmoSApp (Emotional Support App): an entirely offline, smartphone-based conversational app designed to provide mental health and emotional support. EmoSApp leverages a language model, specifically the LLaMA-3.2-1B-Instruct, which is fine-tuned and quantized on a custom-curated ``Knowledge Dataset'' comprising 14,582 mental health QA pairs along with multi-turn conversational data, enabling robust domain expertise and fully on-device inference on resource-constrained smartphones. Through qualitative evaluation with students and mental health professionals, we demonstrate that EmoSApp has the ability to respond coherently and empathetically, provide relevant suggestions to user's mental health problems, and maintain interactive dialogue. Additionally, quantitative evaluations on nine commonsense and reasoning benchmarks, along with two mental health specific datasets, demonstrate EmoSApp's effectiveness in low-resource settings. By prioritizing on-device deployment and specialized domain-specific adaptation, EmoSApp serves as a blueprint for future innovations in portable, secure, and highly tailored AI-driven mental health support.

Figures (5)

• Figure 1: Example responses generated by LLaMA-3.2-1B-Instruct and our proposed EmoSApp. LLaMA model (pink response box) often give generalized and verbose responses, resulting in an AI-generated format. In contrast, EmoSApp (green response box) demonstrates stronger empathy and deeper conversational exploration, effectively providing emotional and mental health support to users.
• Figure 2: An illustration comparing a single weight update in standard Full fine-tuning (left) versus LoRA fine-tuning (right). In LoRA, the low-rank matrices $\mathbf{A}$ and $\mathbf{B}$ approximate the full-rank weight update (highlighted in blue), reducing the number of trainable parameters.
• Figure 3: LLM-based evaluation ratings for Model-A and Model-B across all qualitative metrics. Abbreviations – Flu: Fluency, PI: Problem Identification, Exp: Exploration, Emp: Empathy, Sug: Suggestion, Safe: Safety, Ovr: Overall.
• Figure A.1: Safety demonstration in EmoSApp
• Figure A.2: EmoSApp on a smartphone illustrating: (a) Model loaded state and (b) Offline chat interface.